lustre/mdc/mdc_locks.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * Copyright (C) 2001-2003 Cluster File Systems, Inc.
   5  *
   6  *   This file is part of Lustre, http://www.sf.net/projects/lustre/
   7  *
   8  *   Lustre is free software; you can redistribute it and/or
   9  *   modify it under the terms of version 2 of the GNU General Public
  10  *   License as published by the Free Software Foundation.
  11  *
  12  *   Lustre is distributed in the hope that it will be useful,
  13  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15  *   GNU General Public License for more details.
  16  *
  17  *   You should have received a copy of the GNU General Public License
  18  *   along with Lustre; if not, write to the Free Software
  19  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  20  */
  21
  22 #ifndef EXPORT_SYMTAB
  23 # define EXPORT_SYMTAB
  24 #endif
  25 #define DEBUG_SUBSYSTEM S_MDC
  26
  27 #ifdef __KERNEL__
  28 # include <linux/module.h>
  29 # include <linux/pagemap.h>
  30 # include <linux/miscdevice.h>
  31 # include <linux/init.h>
  32 #else
  33 # include <liblustre.h>
  34 #endif
  35
  36 #include <linux/obd_class.h>
  37 #include <linux/lustre_mds.h>
  38 #include <linux/lustre_dlm.h>
  39 #include <linux/lustre_sec.h>
  40 #include <linux/lprocfs_status.h>
  41 #include <linux/lustre_acl.h>
  42 #include <linux/lustre_lite.h>
  43 #include "mdc_internal.h"
  44
  45 int it_disposition(struct lookup_intent *it, int flag)
  46 {
  47         return LUSTRE_IT(it)->it_disposition & flag;
  48 }
  49 EXPORT_SYMBOL(it_disposition);
  50
  51 void it_set_disposition(struct lookup_intent *it, int flag)
  52 {
  53         LUSTRE_IT(it)->it_disposition |= flag;
  54 }
  55 EXPORT_SYMBOL(it_set_disposition);
  56
  57 static void mdc_id2mdc_data(struct mdc_op_data *data,
  58                             struct lustre_id *f1,
  59                             struct lustre_id *f2,
  60                             const char *name,
  61                             int namelen, int mode)
  62 {
  63         LASSERT(data);
  64         LASSERT(f1);
  65
  66         data->id1 = *f1;
  67         if (f2)
  68                 data->id2 = *f2;
  69
  70         data->valid = 0;
  71         data->name = name;
  72         data->namelen = namelen;
  73         data->create_mode = mode;
  74         data->mod_time = LTIME_S(CURRENT_TIME);
  75 }
  76
  77 static int it_to_lock_mode(struct lookup_intent *it)
  78 {
  79         /* CREAT needs to be tested before open (both could be set) */
  80         if (it->it_op & IT_CREAT)
  81                 return LCK_PW;
  82         else if (it->it_op & (IT_READDIR | IT_GETATTR | IT_OPEN | IT_LOOKUP |
  83                               IT_CHDIR))
  84                 return LCK_PR;
  85
  86         LBUG();
  87         RETURN(-EINVAL);
  88 }
  89
  90 int it_open_error(int phase, struct lookup_intent *it)
  91 {
  92         if (it_disposition(it, DISP_OPEN_OPEN)) {
  93                 if (phase == DISP_OPEN_OPEN)
  94                         return LUSTRE_IT(it)->it_status;
  95                 else
  96                         return 0;
  97         }
  98
  99         if (it_disposition(it, DISP_OPEN_CREATE)) {
 100                 if (phase == DISP_OPEN_CREATE)
 101                         return LUSTRE_IT(it)->it_status;
 102                 else
 103                         return 0;
 104         }
 105
 106         if (it_disposition(it, DISP_LOOKUP_EXECD)) {
 107                 if (phase == DISP_LOOKUP_EXECD)
 108                         return LUSTRE_IT(it)->it_status;
 109                 else
 110                         return 0;
 111         }
 112
 113         if (it_disposition(it, DISP_IT_EXECD)) {
 114                 if (phase == DISP_IT_EXECD)
 115                         return LUSTRE_IT(it)->it_status;
 116                 else
 117                         return 0;
 118         }
 119         CERROR("it disp: %X, status: %d\n", LUSTRE_IT(it)->it_disposition,
 120                LUSTRE_IT(it)->it_status);
 121         LBUG();
 122         return 0;
 123 }
 124 EXPORT_SYMBOL(it_open_error);
 125
 126 /* this must be called on a lockh that is known to have a referenced lock */
 127 int mdc_set_lock_data(struct obd_export *exp, __u64 *l, void *data)
 128 {
 129         struct ldlm_lock *lock;
 130         struct lustre_handle *lockh = (struct lustre_handle *)l;
 131         ENTRY;
 132
 133         if (!*l) {
 134                 EXIT;
 135                 return 0;
 136         }
 137
 138         lock = ldlm_handle2lock(lockh);
 139
 140         LASSERT(lock != NULL);
 141         lock_res_and_lock(lock);
 142 #ifdef __KERNEL__
 143         if (lock->l_ast_data && lock->l_ast_data != data) {
 144                 struct ll_inode_info *lli = ll_i2info(data);
 145                 struct inode *new_inode = data;
 146                 LASSERT(lli->lli_inode_magic == LLI_INODE_MAGIC);
 147                 struct inode *old_inode = lock->l_ast_data;
 148                 if (!(old_inode->i_state & I_FREEING)) {
 149                         CERROR("Found existing inode %p/%lu/%u state %lu "
 150                                "in lock: setting data to %p/%lu/%u\n",
 151                                old_inode, old_inode->i_ino,
 152                                old_inode->i_generation, old_inode->i_state,
 153                                new_inode, new_inode->i_ino,
 154                                new_inode->i_generation);
 155                         unlock_res_and_lock(lock);
 156                         LBUG();
 157                 }
 158         }
 159 #endif
 160         lock->l_ast_data = data;
 161         unlock_res_and_lock(lock);
 162         LDLM_LOCK_PUT(lock);
 163
 164         EXIT;
 165         return 0;
 166 }
 167 EXPORT_SYMBOL(mdc_set_lock_data);
 168
 169 int mdc_change_cbdata(struct obd_export *exp, struct lustre_id *id,
 170                       ldlm_iterator_t it, void *data)
 171 {
 172         struct ldlm_res_id res_id = { .name = {0} };
 173         ENTRY;
 174
 175         res_id.name[0] = id_fid(id);
 176         res_id.name[1] = id_group(id);
 177
 178         ldlm_change_cbdata(class_exp2obd(exp)->obd_namespace,
 179                            &res_id, it, data);
 180
 181         EXIT;
 182         return 0;
 183 }
 184
 185 static inline void
 186 mdc_clear_replay_flag(struct ptlrpc_request *req, int rc)
 187 {
 188         /* Don't hold error requests for replay. */
 189         if (req->rq_replay) {
 190                 unsigned long irqflags;
 191                 spin_lock_irqsave(&req->rq_lock, irqflags);
 192                 req->rq_replay = 0;
 193                 spin_unlock_irqrestore(&req->rq_lock, irqflags);
 194         }
 195         if (rc && req->rq_transno != 0) {
 196                 DEBUG_REQ(D_ERROR, req, "transno returned on error rc %d", rc);
 197                 LBUG();
 198         }
 199 }
 200
 201 /* We always reserve enough space in the reply packet for a stripe MD, because
 202  * we don't know in advance the file type. */
 203 int mdc_enqueue(struct obd_export *exp,
 204                 int lock_type,
 205                 struct lookup_intent *it,
 206                 int lock_mode,
 207                 struct mdc_op_data *data,
 208                 struct lustre_handle *lockh,
 209                 void *lmm,
 210                 int lmmsize,
 211                 ldlm_completion_callback cb_completion,
 212                 ldlm_blocking_callback cb_blocking,
 213                 void *cb_data)
 214 {
 215         struct ptlrpc_request *req;
 216         struct ldlm_res_id res_id = {
 217                 .name = {id_fid(&data->id1), id_group(&data->id1)}
 218         };
 219         struct obd_device *obddev = class_exp2obd(exp);
 220         ldlm_policy_data_t policy = { .l_inodebits = { MDS_INODELOCK_LOOKUP } };
 221         struct ldlm_intent *lit;
 222         struct ldlm_request *lockreq;
 223         int reqsize[6] = {[MDS_REQ_SECDESC_OFF] = 0,
 224                           [MDS_REQ_INTENT_LOCKREQ_OFF] = sizeof(*lockreq),
 225                           [MDS_REQ_INTENT_IT_OFF] = sizeof(*lit)};
 226         int repsize[5] = {sizeof(struct ldlm_reply),
 227                           sizeof(struct mds_body),
 228                           obddev->u.cli.cl_max_mds_easize};
 229         int req_buffers = 3, reply_buffers = 0;
 230         int rc, flags = LDLM_FL_HAS_INTENT;
 231         struct ldlm_reply *dlm_rep = NULL;
 232         void *eadata;
 233         unsigned long irqflags;
 234         ENTRY;
 235
 236 //        LDLM_DEBUG_NOLOCK("mdsintent=%s,name=%s,dir=%lu",
 237 //                          ldlm_it2str(it->it_op), it_name, it_inode->i_ino);
 238
 239         reqsize[0] = lustre_secdesc_size();
 240
 241         if (it->it_op & IT_OPEN) {
 242                 it->it_create_mode |= S_IFREG;
 243                 it->it_create_mode &= ~current->fs->umask;
 244
 245                 reqsize[req_buffers++] = sizeof(struct mds_rec_create);
 246                 reqsize[req_buffers++] = data->namelen + 1;
 247                 reqsize[req_buffers++] = obddev->u.cli.cl_max_mds_easize;
 248
 249                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
 250                                       LDLM_ENQUEUE, req_buffers, reqsize, NULL);
 251                 if (!req)
 252                         RETURN(-ENOMEM);
 253
 254                 spin_lock_irqsave (&req->rq_lock, irqflags);
 255                 req->rq_replay = 1;
 256                 spin_unlock_irqrestore (&req->rq_lock, irqflags);
 257
 258                 /* pack the intent */
 259                 lit = lustre_msg_buf(req->rq_reqmsg, MDS_REQ_INTENT_IT_OFF,
 260                                      sizeof (*lit));
 261                 lit->opc = (__u64)it->it_op;
 262
 263                 /* pack the intended request */
 264                 mdc_open_pack(req->rq_reqmsg, MDS_REQ_INTENT_REC_OFF, data,
 265                               it->it_create_mode, 0, it->it_flags, lmm, lmmsize);
 266                 /* get ready for the reply */
 267                 repsize[3] = 4;
 268                 repsize[4] = xattr_acl_size(LL_ACL_MAX_ENTRIES);
 269                 reply_buffers = 5;
 270                 req->rq_replen = lustre_msg_size(5, repsize);
 271         } else if (it->it_op & (IT_GETATTR | IT_LOOKUP | IT_CHDIR)) {
 272                 __u64 valid = data->valid | OBD_MD_FLNOTOBD | OBD_MD_FLEASIZE |
 273                               OBD_MD_FLACL;
 274
 275                 /* we don't expect xattr retrieve could reach here */
 276                 LASSERT(!(valid & (OBD_MD_FLXATTR | OBD_MD_FLXATTRLIST)));
 277
 278                 reqsize[req_buffers++] = sizeof(struct mds_body);
 279                 reqsize[req_buffers++] = data->namelen + 1;
 280
 281                 if (it->it_op & IT_GETATTR)
 282                         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
 283
 284                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
 285                                       LDLM_ENQUEUE, req_buffers, reqsize, NULL);
 286
 287                 if (!req)
 288                         RETURN(-ENOMEM);
 289
 290                 /* pack the intent */
 291                 lit = lustre_msg_buf(req->rq_reqmsg, MDS_REQ_INTENT_IT_OFF,
 292                                      sizeof (*lit));
 293                 lit->opc = (__u64)it->it_op;
 294
 295                 /* pack the intended request */
 296                 mdc_getattr_pack(req->rq_reqmsg, MDS_REQ_INTENT_REC_OFF,
 297                                  valid, it->it_flags, data);
 298
 299                 /* get ready for the reply */
 300                 repsize[3] = 4;
 301                 repsize[4] = xattr_acl_size(LL_ACL_MAX_ENTRIES);
 302                 reply_buffers = 5;
 303                 req->rq_replen = lustre_msg_size(5, repsize);
 304         } else if (it->it_op == IT_READDIR) {
 305                 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
 306                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
 307                                       LDLM_ENQUEUE, 2, reqsize, NULL);
 308
 309                 if (!req)
 310                         RETURN(-ENOMEM);
 311                 /* get ready for the reply */
 312                 reply_buffers = 1;
 313                 req->rq_replen = lustre_msg_size(1, repsize);
 314         } else if (it->it_op == IT_UNLINK) {
 315                 reqsize[req_buffers++] = sizeof(struct mds_body);
 316                 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
 317                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
 318                                       LDLM_ENQUEUE, req_buffers, reqsize, NULL);
 319                 if (!req)
 320                         RETURN(-ENOMEM);
 321
 322                 /* pack the intended request */
 323                 mdc_getattr_pack(req->rq_reqmsg, MDS_REQ_INTENT_REC_OFF,
 324                                  0, 0, data);
 325
 326                 /* pack the intent */
 327                 lit = lustre_msg_buf(req->rq_reqmsg, MDS_REQ_INTENT_IT_OFF,
 328                                      sizeof (*lit));
 329                 lit->opc = (__u64)it->it_op;
 330
 331                 /* get ready for the reply */
 332                 reply_buffers = 3;
 333                 req->rq_replen = lustre_msg_size(3, repsize);
 334         } else {
 335                 LBUG();
 336                 RETURN(-EINVAL);
 337         }
 338
 339         lustre_pack_secdesc(req, reqsize[0]);
 340
 341         mdc_get_rpc_lock(obddev->u.cli.cl_rpc_lock, it);
 342         rc = ldlm_cli_enqueue(exp, req, obddev->obd_namespace, res_id,
 343                               lock_type, &policy, lock_mode, &flags,cb_blocking,
 344                               cb_completion, NULL, cb_data, NULL, 0, NULL,
 345                               lockh);
 346         mdc_put_rpc_lock(obddev->u.cli.cl_rpc_lock, it);
 347
 348         /* Similarly, if we're going to replay this request, we don't want to
 349          * actually get a lock, just perform the intent. */
 350         if (req->rq_transno || req->rq_replay) {
 351                 lockreq = lustre_msg_buf(req->rq_reqmsg,
 352                                          MDS_REQ_INTENT_LOCKREQ_OFF,
 353                                          sizeof (*lockreq));
 354                 lockreq->lock_flags |= LDLM_FL_INTENT_ONLY;
 355         }
 356
 357         /* This can go when we're sure that this can never happen */
 358         LASSERT(rc != -ENOENT);
 359         /* We need dlm_rep to be assigned this early, to check lock mode of
 360            returned lock from request to avoid possible race with lock
 361            conversion */
 362         if (rc == ELDLM_LOCK_ABORTED || !rc) {
 363                 dlm_rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*dlm_rep));
 364                 LASSERT(dlm_rep != NULL);   /* checked by ldlm_cli_enqueue() */
 365         }
 366         if (rc == ELDLM_LOCK_ABORTED) {
 367                 lock_mode = 0;
 368                 memset(lockh, 0, sizeof(*lockh));
 369                 rc = 0;
 370         } else if (rc != 0) {
 371                 CERROR("ldlm_cli_enqueue: %d\n", rc);
 372                 LASSERTF(rc < 0, "rc = %d\n", rc);
 373                 mdc_clear_replay_flag(req, rc);
 374                 ptlrpc_req_finished(req);
 375                 RETURN(rc);
 376         } else { /* rc = 0 */
 377                 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
 378                 LASSERT(lock);
 379
 380                 /* If the server gave us back a different lock mode, we should
 381                  * fix up our variables. */
 382                 if (dlm_rep->lock_desc.l_req_mode != lock_mode) {
 383                         ldlm_lock_addref(lockh, dlm_rep->lock_desc.l_req_mode);
 384                         ldlm_lock_decref(lockh, lock_mode);
 385                         lock_mode = dlm_rep->lock_desc.l_req_mode;
 386                 }
 387
 388                 ldlm_lock_allow_match(lock);
 389                 LDLM_LOCK_PUT(lock);
 390         }
 391
 392         LASSERT_REPSWABBED(req, 0);         /* swabbed by ldlm_cli_enqueue() */
 393
 394         LUSTRE_IT(it)->it_disposition = (int) dlm_rep->lock_policy_res1;
 395         LUSTRE_IT(it)->it_status = (int) dlm_rep->lock_policy_res2;
 396         LUSTRE_IT(it)->it_lock_mode = lock_mode;
 397         LUSTRE_IT(it)->it_data = req;
 398
 399         if (LUSTRE_IT(it)->it_status < 0 && req->rq_replay)
 400                 mdc_clear_replay_flag(req, LUSTRE_IT(it)->it_status);
 401
 402         DEBUG_REQ(D_DLMTRACE, req, "disposition: %x, status: %d",
 403                   LUSTRE_IT(it)->it_disposition, LUSTRE_IT(it)->it_status);
 404
 405         /* We know what to expect, so we do any byte flipping required here */
 406         LASSERT(reply_buffers == 5 || reply_buffers == 4 ||
 407                 reply_buffers == 3 || reply_buffers == 1);
 408         if (reply_buffers >= 3) {
 409                 struct mds_body *body;
 410
 411                 body = lustre_swab_repbuf(req, 1, sizeof (*body),
 412                                           lustre_swab_mds_body);
 413                 if (body == NULL) {
 414                         CERROR ("Can't swab mds_body\n");
 415                         RETURN (-EPROTO);
 416                 }
 417
 418                 if ((body->valid & OBD_MD_FLEASIZE) != 0) {
 419                         /* The eadata is opaque; just check that it is there.
 420                          * Eventually, obd_unpackmd() will check the contents */
 421                         eadata = lustre_swab_repbuf(req, 2, body->eadatasize,
 422                                                     NULL);
 423                         if (eadata == NULL) {
 424                                 CERROR ("Missing/short eadata\n");
 425                                 RETURN (-EPROTO);
 426                         }
 427                         if (it->it_op & IT_OPEN) {
 428                                 void *replayea;
 429
 430                                 replayea = lustre_msg_buf(req->rq_reqmsg,
 431                                                           MDS_REQ_INTENT_REC_OFF + 2,
 432                                                           body->eadatasize);
 433                                 LASSERT(replayea);
 434                                 memcpy(replayea, eadata, body->eadatasize);
 435
 436                                 LASSERT(req->rq_reqmsg->bufcount == 6);
 437                                 req->rq_reqmsg->buflens[5] = body->eadatasize;
 438                                 /* If this isn't the last buffer, we might
 439                                  * have to shift other data around. */
 440                         }
 441                 }
 442         }
 443
 444         RETURN(rc);
 445 }
 446 EXPORT_SYMBOL(mdc_enqueue);
 447
 448 /*
 449  * This long block is all about fixing up the lock and request state
 450  * so that it is correct as of the moment _before_ the operation was
 451  * applied; that way, the VFS will think that everything is normal and
 452  * call Lustre's regular VFS methods.
 453  *
 454  * If we're performing a creation, that means that unless the creation
 455  * failed with EEXIST, we should fake up a negative dentry.
 456  *
 457  * For everything else, we want to lookup to succeed.
 458  *
 459  * One additional note: if CREATE or OPEN succeeded, we add an extra
 460  * reference to the request because we need to keep it around until
 461  * ll_create/ll_open gets called.
 462  *
 463  * The server will return to us, in it_disposition, an indication of
 464  * exactly what d.lustre->it_status refers to.
 465  *
 466  * If DISP_OPEN_OPEN is set, then d.lustre->it_status refers to the open() call,
 467  * otherwise if DISP_OPEN_CREATE is set, then it status is the
 468  * creation failure mode.  In either case, one of DISP_LOOKUP_NEG or
 469  * DISP_LOOKUP_POS will be set, indicating whether the child lookup
 470  * was successful.
 471  *
 472  * Else, if DISP_LOOKUP_EXECD then d.lustre->it_status is the rc of the
 473  * child lookup.
 474  */
 475 int mdc_intent_lock(struct obd_export *exp, struct lustre_id *pid,
 476                     const char *name, int len, void *lmm, int lmmsize,
 477                     struct lustre_id *cid, struct lookup_intent *it,
 478                     int lookup_flags, struct ptlrpc_request **reqp,
 479                     ldlm_blocking_callback cb_blocking)
 480 {
 481         struct lustre_handle lockh;
 482         struct ptlrpc_request *request;
 483         struct mds_body *mds_body;
 484         struct lustre_handle old_lock;
 485         struct ldlm_lock *lock;
 486         int rc = 0;
 487         ENTRY;
 488         LASSERT(it);
 489
 490         CDEBUG(D_DLMTRACE, "name: %*s in obj "DLID4", intent: %s flags %#o\n",
 491                len, name, OLID4(pid), ldlm_it2str(it->it_op), it->it_flags);
 492
 493         if (cid && (it->it_op == IT_LOOKUP || it->it_op == IT_GETATTR ||
 494                     it->it_op == IT_CHDIR)) {
 495                 /* We could just return 1 immediately, but since we should only
 496                  * be called in revalidate_it if we already have a lock, let's
 497                  * verify that. */
 498                 struct ldlm_res_id res_id = {.name = {id_fid(cid),
 499                                                       id_group(cid)}};
 500                 struct lustre_handle lockh;
 501                 ldlm_policy_data_t policy;
 502                 int mode;
 503
 504                 /* For the GETATTR case, ll_revalidate_it issues two separate
 505                    queries - for LOOKUP and for UPDATE lock because it cannot
 506                    check them together - we might have those two bits to be
 507                    present in two separate granted locks */
 508                 policy.l_inodebits.bits = (it->it_op == IT_GETATTR) ?
 509                         MDS_INODELOCK_UPDATE : MDS_INODELOCK_LOOKUP;
 510
 511                 mode = LCK_PR;
 512                 rc = ldlm_lock_match(exp->exp_obd->obd_namespace,
 513                                      LDLM_FL_BLOCK_GRANTED, &res_id,
 514                                      LDLM_IBITS, &policy, mode,
 515                                      &lockh);
 516
 517                 if (!rc) {
 518                         mode = LCK_CR;
 519                         rc = ldlm_lock_match(exp->exp_obd->obd_namespace,
 520                                              LDLM_FL_BLOCK_GRANTED, &res_id,
 521                                              LDLM_IBITS, &policy, mode,
 522                                              &lockh);
 523                 }
 524                 if (!rc) {
 525                         mode = LCK_PW;
 526                         rc = ldlm_lock_match(exp->exp_obd->obd_namespace,
 527                                              LDLM_FL_BLOCK_GRANTED, &res_id,
 528                                              LDLM_IBITS, &policy, mode,
 529                                              &lockh);
 530                 }
 531                 if (!rc) {
 532                         mode = LCK_CW;
 533                         rc = ldlm_lock_match(exp->exp_obd->obd_namespace,
 534                                              LDLM_FL_BLOCK_GRANTED, &res_id,
 535                                              LDLM_IBITS, &policy, mode,
 536                                              &lockh);
 537                 }
 538                 if (rc) {
 539                         if (ptlrpcs_check_cred(exp->exp_obd->u.cli.cl_import)) {
 540                                 /* return immediately if no credential held */
 541                                 ldlm_lock_decref(&lockh, mode);
 542                                 RETURN(-EACCES);
 543                         }
 544                         memcpy(&LUSTRE_IT(it)->it_lock_handle, &lockh,
 545                                sizeof(lockh));
 546                         LUSTRE_IT(it)->it_lock_mode = mode;
 547                 }
 548
 549                 /* Only return failure if it was not GETATTR by cid (from
 550                    inode_revalidate) */
 551                 if (rc || name)
 552                         RETURN(rc);
 553         }
 554
 555         /* lookup_it may be called only after revalidate_it has run, because
 556          * revalidate_it cannot return errors, only zero.  Returning zero causes
 557          * this call to lookup, which *can* return an error.
 558          *
 559          * We only want to execute the request associated with the intent one
 560          * time, however, so don't send the request again.  Instead, skip past
 561          * this and use the request from revalidate.  In this case, revalidate
 562          * never dropped its reference, so the refcounts are all OK */
 563         if (!it_disposition(it, DISP_ENQ_COMPLETE)) {
 564                 struct mdc_op_data *op_data;
 565
 566                 OBD_ALLOC(op_data, sizeof(*op_data));
 567                 if (op_data == NULL)
 568                         RETURN(-ENOMEM);
 569
 570                 mdc_id2mdc_data(op_data, pid, cid, name, len, 0);
 571
 572                 if (name != NULL)
 573                         op_data->valid |= OBD_MD_FID;
 574
 575                 rc = mdc_enqueue(exp, LDLM_IBITS, it, it_to_lock_mode(it),
 576                                  op_data, &lockh, lmm, lmmsize,
 577                                  ldlm_completion_ast, cb_blocking, NULL);
 578                 OBD_FREE(op_data, sizeof(*op_data));
 579                 if (rc < 0)
 580                         RETURN(rc);
 581
 582                 memcpy(&LUSTRE_IT(it)->it_lock_handle, &lockh, sizeof(lockh));
 583         }
 584         request = *reqp = LUSTRE_IT(it)->it_data;
 585         LASSERT(request != NULL);
 586
 587         /* If we're doing an IT_OPEN which did not result in an actual
 588          * successful open, then we need to remove the bit which saves this
 589          * request for unconditional replay.
 590          *
 591          * It's important that we do this first!  Otherwise we might exit the
 592          * function without doing so, and try to replay a failed create (bug
 593          * 3440) */
 594         if (it->it_op & IT_OPEN && request->rq_replay &&
 595             (!it_disposition(it, DISP_OPEN_OPEN) || LUSTRE_IT(it)->it_status != 0))
 596                 mdc_clear_replay_flag(request, LUSTRE_IT(it)->it_status);
 597
 598         if (!it_disposition(it, DISP_IT_EXECD)) {
 599                 /* The server failed before it even started executing the
 600                  * intent, i.e. because it couldn't unpack the request. */
 601                 LASSERT(LUSTRE_IT(it)->it_status != 0);
 602                 RETURN(LUSTRE_IT(it)->it_status);
 603         }
 604         rc = it_open_error(DISP_IT_EXECD, it);
 605         if (rc)
 606                 RETURN(rc);
 607
 608         mds_body = lustre_msg_buf(request->rq_repmsg, 1, sizeof(*mds_body));
 609         LASSERT(mds_body != NULL);      /* mdc_enqueue checked */
 610         LASSERT_REPSWABBED(request, 1); /* mdc_enqueue swabbed */
 611
 612         /* If we were revalidating a fid/name pair, mark the intent in case we
 613          * fail and get called again from lookup */
 614         if (cid != NULL) {
 615                 it_set_disposition(it, DISP_ENQ_COMPLETE);
 616                 /* Also: did we find the same inode? */
 617
 618                 /* we have to compare all the fields but type, because MDS can
 619                  * return fid/mds/ino/gen if inode lives on another MDS -bzzz */
 620                 if (!(lookup_flags & LOOKUP_COBD) && !id_equal(cid, &mds_body->id1))
 621                         RETURN(-ESTALE);
 622         }
 623
 624         rc = it_open_error(DISP_LOOKUP_EXECD, it);
 625         if (rc)
 626                 RETURN(rc);
 627
 628         /*
 629          * keep requests around for the multiple phases of the call this shows
 630          * the DISP_XX must guarantee we make it into the call.
 631          */
 632         if (it_disposition(it, DISP_OPEN_CREATE) &&
 633             !it_open_error(DISP_OPEN_CREATE, it))
 634                 ptlrpc_request_addref(request); /* balanced in ll_create_node */
 635         if (it_disposition(it, DISP_OPEN_OPEN) &&
 636             !it_open_error(DISP_OPEN_OPEN, it))
 637                 ptlrpc_request_addref(request); /* balanced in ll_file_open */
 638
 639         if (it->it_op & IT_CREAT) {
 640                 /* XXX this belongs in ll_create_it */
 641         } else if (it->it_op == IT_OPEN) {
 642                 LASSERT(!it_disposition(it, DISP_OPEN_CREATE));
 643         } else {
 644                 LASSERT(it->it_op & (IT_GETATTR | IT_LOOKUP | IT_CHDIR));
 645         }
 646
 647         /*
 648          * if we already have a matching lock, then cancel the new one. We have
 649          * to set the data here instead of in mdc_enqueue, because we need to
 650          * use the child's inode as the l_ast_data to match, and that's not
 651          * available until intent_finish has performed the iget().)
 652          */
 653         lock = ldlm_handle2lock(&lockh);
 654         if (lock) {
 655                 ldlm_policy_data_t policy = lock->l_policy_data;
 656                 LDLM_DEBUG(lock, "matching against this");
 657                 LDLM_LOCK_PUT(lock);
 658
 659                 LASSERTF(id_fid(&mds_body->id1) == lock->l_resource->lr_name.name[0] &&
 660                          id_group(&mds_body->id1) == lock->l_resource->lr_name.name[1],
 661                          "Invalid lock is returned to client. Lock res_is: %lu/%lu, "
 662                          "response res_id: %lu/%lu.\n",
 663                          (unsigned long)lock->l_resource->lr_name.name[0],
 664                          (unsigned long)lock->l_resource->lr_name.name[1],
 665                          (unsigned long)id_fid(&mds_body->id1),
 666                          (unsigned long)id_group(&mds_body->id1));
 667
 668                 memcpy(&old_lock, &lockh, sizeof(lockh));
 669                 if (ldlm_lock_match(NULL, LDLM_FL_BLOCK_GRANTED, NULL,
 670                                     LDLM_IBITS, &policy, LCK_NL, &old_lock)) {
 671                         ldlm_lock_decref_and_cancel(&lockh,
 672                                                     LUSTRE_IT(it)->it_lock_mode);
 673                         memcpy(&lockh, &old_lock, sizeof(old_lock));
 674                         memcpy(&LUSTRE_IT(it)->it_lock_handle, &lockh,
 675                                sizeof(lockh));
 676                 }
 677         }
 678         CDEBUG(D_DENTRY, "D_IT dentry %*s intent: %s status %d disp %x rc %d\n",
 679                len, name, ldlm_it2str(it->it_op), LUSTRE_IT(it)->it_status,
 680                LUSTRE_IT(it)->it_disposition, rc);
 681
 682         RETURN(rc);
 683 }
 684 EXPORT_SYMBOL(mdc_intent_lock);